融合特征交叉与用户历史行为序列的微地图推荐

Integration of Feature Interaction and User Historical Behavior Sequence for WeMaps Recommendation

针对现有微地图(WeMaps)推荐算法未充分挖掘用户与微地图特征,推荐结果准确率较低的问题,提出融合特征交叉与用户历史行为序列的微地图推荐算法。首先,在用户与地图特征交叉过程中通过引入跳跃连接和多头自注意力机制,让不同特征组合能够自动获得权重,并通过在多个子空间下进行特征交叉获得了更丰富的特征组合信息。其次,在用户历史行为序列中引入了交叉注意力机制,捕捉与候选地图相关的兴趣点。最后,融合特征交叉和用户行为序列模块的输出,获得了综合多个维度的推荐结果。在公开数据集Criteo和自制微地图(WeMaps)数据集上的对数损失值分别为0.4461、0.3797,受试者操作特征曲线下面积值(Area Under the ROC Curve,AUC)分别为0.8052、0.7883。相较于本文对比实验中的二阶特征交叉模型,损失值分别降低了1.7%、14.2%,AUC值提高了0.8%、0.4%。相较于本文对比实验中的高阶特征交叉模型,损失值平均降低了1.3%、2.6%,AUC值平均提高了0.6%,0.2%。较低的损失值和较高的AUC值说明模型进行预测时具有较高的准确性和较好的分类能力。实验结果表明,本文算法不但能为用户提供更为准确的推荐结果,也能使推荐结果具备良好的可解释性。

The existing WeMaps recommendation algorithm cannot fully exploit the features of users and maps,resulting in a low recommendation accuracy.In this study,a WeMaps recommendation algorithm is proposed which combines feature interaction and user history behavior sequence.To begin with,a multi-head attention mechanism and skip connections are introduced in the process of feature interaction between the user and the map.The multi-head attention mechanism allows for interactions between different features in multiple subspaces,resulting in richer feature combinations.The skip connections combine low-level and high-level feature interactions,ensuring model effectiveness and avoiding the occurrence of model degradation.Additionally,a cross-attention mechanism is incorporated into the user's historical behavior sequence to identify the points of interest related to the candidate map.This mechanism effectively captures points of interest within the user's historical behavior sequence that align with the recommended map.It assigns greater attention to the recommended map that corresponds to the user's short-term dynamic interests.Finally,by utilizing a deep neural network,the output results from the feature interaction module and the user's historical behavior sequence module are fused together,resulting in a more accurate recommendation that takes multiple dimensions into consideration.The values of log loss on the public datasets Criteo and WeMaps datasets are 0.4461,and 0.3797,respectively,and the values of area under the receiver operating characteristic curve are 0.8052,and 0.7883,respectively.For convenience of expression,the area under the receiver operating characteristic curve will be referred to as AUC in the following text.Compared to the second-order feature interaction model in the experimental study of this paper,the loss values decreases by 1.7%and 14.2%,respectively,while the AUC values increases by 0.8%and 0.4%,respectively Compared to the high-order feature interaction model in the experimental study of this paper,the average decrease in loss values is 1.3%and 2.6%,respectively,while the average increase in AUC values is 0.6%and 0.2%,respectively.When the model has a low loss value and a high AUC value,it indicates that the model has a small discrepancy between its predicted results on the training data and the actual labels,and the model is capable of effectively distinguishing between positive and negative examples at different thresholds.This is generally considered as a good indicator of model performance,indicating that the model has higher accuracy and better classification ability.Experimental results show that the model proposed in this paper can not only provide users with more accurate recommendation results but also make the recommendation results have good explainability.